A model achieves 99% training accuracy but only 72% validation accuracy. Adding Dropout (p=0.5) to hidden layers reduces training accuracy to 91% but improves validation accuracy to 85%. A junior engineer is alarmed: "Dropout hurt our training accuracy!" Is this a problem?